SmollerTTS Text to speech

A 2.81M-parameter English voice TTS model that runs on your browser’s CPU.

Loading

Loading the voice…

Speech quality vs. model size (rough)

SmollerTTSUnder 10M10–100M100M+
UTMOS ↑
Parameters · log scale
Method & sources
All results
Model results. Select a column heading to sort.
Source

Inside SmollerTTS

From phonemes to a waveform. Select a block to explore.

Swipe the diagram to see all three columns.

Model details

A phoneme-based acoustic model predicts an 80-band mel spectrogram. HiFi-GAN V2 turns it into speech.

Our acoustic weights were trained on LJSpeech. The vocoder uses pretrained LJ_V2 weights.

References & credits ↗
Voice
EnglishSingle speaker
Parameters
2.81M1.88M acoustic + 0.93M vocoder
Audio output
22.05 kHzMono · WAV download
Inference
Browser CPUONNX Runtime · WebAssembly
Weight files
11.34 MBFP32 · ONNX
Browser files
—Includes runtime & pronunciation